一种基于深度学习模型的图像模糊自动分析处理算法

Automatic Analysis and Processing of Image Blur Using Deep Learning

现有的图像去模糊算法不能有效去除图像中存在的局部运动模糊,并且现有的图像局部模糊区域检测算法仅使用低维模糊特征进行图像模糊程度的度量,这会导致模糊区域检测结果出现误检测.针对上述问题,本文将图像局部模糊区域检测和图像去运动模糊两项技术进行有效结合,提出一种基于深度学习的局部运动模糊图像去模糊方法.首先,本文提出一个基于自编码神经网络的深度学习框架,该框架能够准确地标记出输入图像中的局部模糊区域.然后,将这些检测出的局部模糊区域作为遮罩层,仅对这些区域进行去模糊处理,这样就能够在有效去除局部运动模糊的同时不发生图像失真,最终重建出令人满意的去模糊图像.为验证算法的有效性,对算法结果进行主、客观评价并与现存算法进行比较.精确度-查全率曲线表明,该算法在相同查全率下较现存算法实现了最高的准确率.同时,该算法能够得到比现存图像去模糊算法更加清晰的重建图像.

As the existing image deblurring methods do not apply to the image degraded by partial motion blur, and the existing partial blur detection approaches only used low-level blur features to measure the blurry degree of an image, the blur regions extracted via these methods usually have misclassification. To address these problems ,we combine the partial blur classification method and image deblurring technique together and present an image deblurring method based on deep learning to eliminate the partial blur. Firstly, we present a deep learning framework with a stacked auto-encoder, which can be used to detect and classify blurred regions in a partially blurred image. Then, we treat these blurred regions as mask layer, and only remove the motion blur in these blurred regions. In this way, we can not only eliminate the partial blur but also reduce image distortion,and finally restore a satisfactory deblurred image. To evaluate the effectiveness of our method, we compare the proposed algorithm with several state-of-the-art techniques both in objective indices and subjective visual experience. Quantitative comparison via precision-recall curve demonstrates that our method achieves the highest precision within almost the entire recall range. Meanwhile, the images restored by our method are sharper than the existing techniques.